3 SAMPLING OF THE VOICE SIGNAL FOR TELEPHONY

Objectives
This practice is intended to enable a good understanding of the sampling concept,
applied for the voice signal in telephony. The rules that have to be considered when
sampling the voice signal in order to allow the reconstruction of the original signal from
samples are explained. The original, sampled and reconstructed signals are analyzed both
in time and frequency domain, in order to evaluate the quality of the reconstructed signal.
The effects of aliasing errors and of interference on the reconstructed signal are studied.
The Feedback 58-001 USB system together with the TDM/PCM principles work board
are used for the practical study of sampling. The signals can be observed both using the
virtual instruments in the Discovery IMS environment and using a physical oscilloscope.

3.1 Theoretical background

The voice signal for telephony is an analog signal, continuous both in time and
level domains. The transmission of the signal over long distances in analog format has
multiple drawbacks, of which, amongst the most important are the cumulative effect of
noise and interference and the presence of non-linear distortions. Even if the noise and
interference levels are very low, the receiver cannot separate the original signal from
noise and interference. Therefore, the received signal in an analog communication
system is only an approximation of the original signal (not identical). On the other hand,
in a digital communication system, the received signal can be reconstructed as an
identical copy of the original signal, as long as the levels of noise and interference are
kept within acceptable limits. It is therefore desirable to transmit the signal in digital
format.
A digital communication system works with signals represented by discrete
levels (values), at discrete moments in time. In order to enable the digital transmission
for an analog signal, some additional steps are required prior to transmission and after
receiving the signal, as shown in Fig. 3.1. On the transmit side, the continuity of the
analog signal must be broken, both in the temporal domain (through sampling) and in
the level domain (through quantization). After the quantization and ADC each sample is
represented by a binary code, which is transmitted on the communication channel using
a specific line code. On the receive side the DAC and the reconstruction low-pass filter

Quantization 010010 Reconstruction

Sampling DAC
and ADC LPF
Discrete
levels
Discrete pulses (in time domain)

Fig. 3.1. Digital transmission of an analog signal

(LPF) restore the continuity of the signal in the 2 domains offering at the output an analog
signal. The sampling and its inverse operation (LPF for reconstruction) are discussed in
this chapter.
26 Sampling of the voice signal for telephony – 3

Sampling
The sampling is a process through which an analog signal, continuous in the time
domain, is represented by a series of discrete values (the samples of the signal). The
samples are taken at equally spaced moments in time by closing/opening a switch (as
shown in Fig. 3.2). The sampling process is a basic operation for any application of
storage or transmission of analog signals in discrete/digital format.
The periodical opening/closing of the sampling switch can be implemented by
controlling the switch actions through a periodical signal containing rectangular pulses.
The result at the output of the sampling circuit is a series of periodical pulses, having the
amplitudes modulated by the input analog signal, as shown in Fig. 3.3.
Using a mathematical approach, the sampling process can be seen as a product
between the analog input signal and the pulse train signal. The circuit in Fig. 3.2 is similar
to the simple balanced AMDSB-SC modulator from Fig. 1.6.a), therefore the sampling
process can also be seen as a pulse amplitude modulation (PAM) process.
While for AM the signal which controls the switch has a duty cycle of 50% (it is
either a sine or clock signal), for sampling the pulses in the control signal usually have a
duty cycle (much) lower than 50%. The sampling circuit actually copies the input signal
to the output when the switch is closed (when the pulse is active) and gives 0 at the output
in the rest of the time. Given the objective for digital transmission – to convert each
sample to a numerical value – it is desirable not to have significant variations in the level
of the input signal while the sampling switch is closed. The usage of narrow pulse widths
reduces the time window in which the input signal is copied at the output resulting in
limited variations of the signal level during this short time interval.
A periodical rectangular signal can be written using the Fourier series expansion as:

x t a0 an cos 2 nf s t (3.1)
n 1

Equation (3.1) suggests that the rectangular signal x t – with pulse frequency
fs – is obtained by adding an infinity of cosine waves (with frequencies at integer

pulses(t)

analog(t) sampled(t)

Fig. 3.2. Sampling circuit

 3.1 – Theoretical background 27

analog
(input) t

sampling
pulses
t

PAM
(output) t

Fig. 3.3. Time representation of the signals involved in the sampling process

f a
multiples of the fundamental frequency, s , and different amplitudes, n ) to a DC
a a
component, 0. The amplitudes n of the cosine waves can be calculated using the
equation [rs1]:

2
an Asin n d (3.2)
n
In equation (3.2), A represents the amplitude of the pulses in the rectangular signal and
d is the duty cycle of the pulses. Given that the sine function crosses 0 at each integer

multiple of , the term sin n d in equation (3.2) will be 0 if nd ,

causing the corresponding harmonics in equation (3.1) to have 0 amplitude. For example,
a square wave (clock signal) – for which d 12 – will have all the even harmonics in
equation (3.1) with 0 amplitude; a rectangular signal with duty cycle d 13 will have all
the integer multiples of 3 harmonics with 0 amplitude and so on.
Considering the Fourier series representation of the pulse train signal in equation
(3.1), the spectral representation of such a signal looks like in Fig. 3.4. The spectral
representation of a pulse train signal has non-zero values only at 0 (DC component), at
fs and at the harmonics of fs .
Given the above mentioned similarity of the sampling circuit with the
AMDSBSC modulator, the result of sampling, in the frequency domain, is a pair of side-
lobes around each of the non-zero spectral components of the sampling pulse train signal.
The signals used in transmissions are band limited signals – all their spectral components
f
above a maximum frequency max are 0). Therefore, the side-lobes around a given
f
harmonic will extend with the amount of max above and below that harmonic.
28 Sampling of the voice signal for telephony – 3

t
T=1/fs

...
0 fs 2fs 3fs 4fs 5fs f

Fig. 3.4. Time and spectral representations of the sampling pulse-train signal

Fig. 3.5 shows the frequency aligned spectral representations of a bandlimited

analog signal, the sampling pulses and the sampled signal (the result of sampling). The
spectrum of the sampled voice signal (Fig. 3.5 – bottom) consits of:
f
• a copy of the spectrum of the original analog signal (between 0 – max ),
• side-lobes (replica of the original signal’s spectrum) around the fundamental
frequency of the sampling pulses (red) and
• side-lobes around each harmonic of the fundamental frequency (blue, magenta
etc.).
Apparently, looking at the time representation of the sampled signal in Fig. 3.3
one may say that parts of the original signal (between samples) are lost. On the other
hand, the analysis of the spectral representation of the sampled signal in Fig. 3.5 reveals
that the sampled signal contains all the original (base-band) information and even more:
an infinity of copies of the original spectrum centered around the fundamental frequency
of the sampling pulses and around all the harmonics of this frequency.
As mentioned before, sampling is one of the operations necessary to prepare the
analog voice signal for digital transmission, but on the receive side, the operation must
be reversed in order to restore the signal’s continuity in time domain. The choice of the
sampling frequency (the frequency of the sampling pulses) is critical for the posibility

-fmax 0 fmax f

0 fs 2fs 3f s f

0 fmax fs 2fs 3f s f
fs - fmax fs + fmax
 3.1 – Theoretical background 29

Fig. 3.5. Spectra of the analog signal (top), sampling pulses (middle) and sampled signal (bottom)

to restore the original analog signal from its samples. The sampling theorem (also known
as the Nyquist theorem or the Shannon sampling theroem) [rs1] establishes the rule for
the proper choice of the sampling frequency.
The sampling theorem:
f
A bandlimited analog signal, with the highest frequency max , is uniquely
determined by its samples and can be reconstructed from them using an ideal low-pass
f
filter, if the sampling frequency, s , is higher than the double of the maximum frequency
in the spectrum of the original analog signal:
fs 2 fmax. (3.3)
In order to reject the high frequency side-lobes and keep only the copy of the original
f
spectrum, the cut-off frequency of the reconstruction low-pass filter, c, must fulfill the
following criterion:

fmax fc fs fmax (3.4)

Reconstruction
By analizing the spectrum of the sampled signal – when the Nyquist criterion is
met – the reconstruction process on the receive side appearss straightforward. It is
obvious that the use of a low-pass filter with a properly chosen cut-off frequency may
remove all the higher frequency side-lobes. After this removal, the spectrum of the signal
contains only the copy of the original signal (possibly attenuated), which can be returned
to the original form by a fixed gain amplification. Fig. 3.6 presents the spectra of the
sampled signal (top) and reconstructed signal (bottom) together with the frequency
response of the reconstruction LPF (middle).

0 fmax fs 2fs 3fs f

fs - fmax fs + fmax

0 fc f

0 fmax f

Fig. 3.6. Reconstruction of the analog signal from its samples – spectral representations

Given that f f
s 2 max , a good choice for the cut-off frequency of the ideal
30 Sampling of the voice signal for telephony – 3

LPF is half the value of the sampling frequency. However, in practical applications ideal
LPFs are not available. The real LPFs have a non ideal frequency response characteristic.
Therefore a guard frequency band must be ensured between the spectrum of the original
f
signal (which extends to the right up to max f) and the closest higher frequency side-lobe
(which extends to the left up to fs fmax ). For practical aplications, it is recommended to
choose the sampling frequency slightly higher than twice the maximum frequency in the
spectrum of the original signal in order to allow the non-ideal filter to separate the useful
components from the higher frequency replicas:

fs 1.1 1.3 2 fmax. (3.5)

Aliasing
If the Nyquist criterion in equation (3.3) is not fulfilled, spectral aliasing may
occur (higher frequency side-lobes overlap the useful frequency components in the
baseband). If spectral aliasing occurs, it is impossible to separate the useful base-band
components of the original signal from the overlapping side-lobe(s), resulting in
reconstruction (aliasing) errors.
Fig. 3.7 presents the result of aliasing on the spectrum of the sampled signal.
Similar to Fig. 3.5, the top and middle plots represent the spectra of the original signal
and of the sampling pulses, but this time the sampling frequency is lower and does not
fulfill the Nyquist criterion. The result of sampling, shown in the bottom plot, reveals

-fmax 0 fmax f

fs < 2fmax

0 fs 2fs 3fs f

aliasing

0 fmax fs 2fs 3fs f

fs - fmax fs + fmax

Fig. 3.7. Aliasing errors in the spectrum of the sampled signal

overlapping between adjacent side-lobes. In the base-band of the original signal, an

overlapping occurs between fs fmax (which is now lower than fmax ) and fmax . In this
case, the reconstruction using a low-pass filter is not possible, regardless of the value of
the cut-off frequency. Once the spectral components of the useful original signal and the
unwanted side-lobe are combined together, there is no way to separate them and any
attempt of reconstruction leads to a reconstructed signal with aliasing errors.
 3.1 – Theoretical background 31

Sampling of the voice signal

The voice signal transmitted in telephony is limited in the frequency domain to
the band between 300 Hz and 3.4 kHz. The above mentioned bandwidth is the result of
a compromise between frequency bandwidth and quality. The human voice produces
also frequencies above 3.4 kHz, but the frequencies between 0.3 – 3.4 kHz ensure a
reasonable quality of the signal that enables the recognition of the speaker’s voice and
the understanding of the message.
Considering the Nyquist criterion, and the need to use a non-ideal low-pass filter
for reconstruction, the sampling frequency must exceed the double of the maximum
frequency in the spectrum of the useful signal as recommended in equation (3.5).
According to equation (3.5), the sampling frequency for the voice signal in telephony
must be between 7.48 and 8.84 kHz. The frequency of 8 kHz was chosen and adopted as
worldwide standard sampling frequency for the voice signal transmitted in the digital
telephony: fs 8 kHz. (3.6)
Knowing that the human voice may produce frequencies higher than 4 kHz and
that the microphone may capture sounds coming also from other sources, with even
higher frequencies (possibly inaudible), it is important to use an anti-aliasing low-pass
filter on the transmit side prior to sampling. The role of the anti-aliasing low-pass filter
is to remove all frequencies equal or higher than half the sampling frequency from the
signal that is to be sampled. In the case of the voice signal for telephony, the anti-aliasing
filter must remove all frequencies above 4 kHz. If the anti-aliasing filter is not used, a
f
tone of 6 kHz sampled with s 8 kHz will result in a reconstructed tone of 2 kHz, which
is not the original signal but an aliasing component.
Interference, noise and distortion
Unwanted signals may add to the original useful signal on a transmission chain
in many ways and cause loss of quality. The loss of quality may vary from hardly
noticeable to unacceptable. There are 3 main classes of disturbances which cause loss of
quality in a transmission system: interference, noise and distortion.
Interference is represented by any signal not coming from the input. An example
of interference specific to telephony systems is the crosstalk. In this case signal from one
voice circuit is picked-up (sometimes intelligible) in another voice circuit. This may
happen when the pairs of wires carrying the 2 voice circuits are close to each other (for
example in a cable with multiple pairs). As mentioned in the beginning of the theoretical
part, in an analog transmission, any interference added to the useful signal cannot be
removed and leads to quality loss. In digital transmissions the interference has no effect
up to a given threshold (the decision threshold between neighboring symbols), but if that
threshold is exceeded the errors in the digital signal may have catastrophic consequences,
destroying the useful signal.
Noise is present in all electrical circuits and is random in frequency and
amplitude. An important cause of the noise is the thermal agitation.
Distortion is represented by any change in original signal on the transmission
chain. Unlike interference and noise which are signal independent, distortion only exists
in the presence of the useful signal. Nonlinear distortions caused by the non-ideal
functioning of amplifiers cannot be compensated. Linear distortions caused by variable
attenuation at different frequencies can be compensated with equalizers. The
quantization noise (inherent in digital transmission/storage of analog signals – including
32 Sampling of the voice signal for telephony – 3

voice) can also be considered a form of distortion as it only exists in the presence of the
signal.

3.2 Experimental models

The TDM & PCM principles workboard 58-001-USB (Feedback Instruments)
allows the study of sampling in the context of the transmission of the voice signal for
telephony [rs2]. The practical applications rely on a web-based interface (Discovery
IMS) which presents a block diagram of the studied system with test points to which
virtual instruments (usually oscilloscope) can be connected. For certain applications,
other virtual instruments (e.g. frequency meter, voltmeter etc.) can be present in the block
diagram connected directly to certain test points of interest. The signals can also be
analyzed using real instruments connected to the test points marked on the workboard.
The numbering of the test points on the block diagram in the Discovery IMS and on the
physical workboard is the same.
Depending on the application at hand, some parameters can be adjusted from
tuning knobs on the physical workboard, while other parameters can be changed from
active components in the block diagram available in the web interface of Discovery IMS.
The tuning knobs effective in a given application are displayed in the lower-left corner
of the system block diagram in Discovery IMS (see Fig. 3.8). Before starting a practical
application all relevant tuning knobs on the workboard must be adjusted to the initial
position indicated in the lower-left corner of the block diagram.

Fig. 3.8. Block diagram of the system for the study of basic sampling
 33
3.2 – Experimental models
For the study of sampling of the voice signal for telephony, the Discovery IMS
provides 4 practical application models: Basic sampling, Aliasing errors, Interference
and Spectrum analysis.
Basic sampling model
The model used for the study of basic sampling, presented in Fig. 3.8, consists
of a sampling block and a low-pass reconstruction filter. The sampling block takes 2
input signals: the signal to sample (from the oscillator Osc 1) and a sampling clock.
The frequency and amplitude of the sine wave produced by the oscillator Osc 1
can be adjusted from the tuning knobs shown in Fig. 3.8. The 2 knobs for frequency
adjustment (‘Low’ and ‘High’) allow the user to set frequency values in the range 200
Hz – 12 kHz. The frequency meter below the Osc 1 block displays the current value of
the frequency of the sine wave produced by the oscillator Osc 1. The 2 knobs for
amplitude setting (‘Output Coarse’ and ‘Fine’) allow amplitudes of the sine wave to be
varied in the range 0 – 2.1 V.
The frequency of the sampling pulses is fixed at the standard sampling rate for
the telephony voice – 8 kHz. The only parameter that can be adjusted for the sampling
pulses is the duty cycle – adjustable through an active button in the block diagram –
Select Sample Time. The available values for the duty cycle are 1/2, 1/4 and 1/8.
The low-pass filter at the output of the block diagram reconstructs the signal
from samples and has no adjustable parameters.
The virtual instrument (oscilloscope) can be connected to any of the following
test points:
TP1 – input signal, TP5 – sampling pulses,
TP7 – sampled signal, TP8 – reconstructed signal.
The signals can also be observed using a physical oscilloscope at the test points with the
same numbers, on the physical workboard. Given that the virtual instruments in the
Discovery IMS are not very precise – especially for measuring amplitudes – it is
recommended to use the physical oscilloscope where precision measurements are
required.
Aliasing errors model
The model used for the study of aliasing errors, presented in Fig. 3.9, is similar
to that used for the basic sampling principles, with the only noticeable difference that the
duty cycle of the sampling pulses can no longer be adjusted (it is set to the default value
of 1/2). The frequency and amplitude of the input signal can still be adjusted from the
tuning knobs on the workboard. The available test points are the same as in the case of
the basic sampling model.
Interference model
The model used for the study of the effects of interferences, presented in Fig.
3.10, keeps the main elements from the model for the study of basic sampling and adds
a source of interference which inserts a disturbing signal between the sampling block
34 Sampling of the voice signal for telephony – 3
and the reconstruction low-pass filter. A second frequency meter is connected at the
output of the system (TP8).

Fig. 3.9. Block diagram of the system for the study of aliasing errors

Fig. 3.10. Block diagram of the system for the study of the effect of interference
 35
Fig. 3.11. Block diagram of the spectrum analysis model
3.3 – Practice and exercises
A new test point is available: TP4 – where the interference signal can be viewed.
In this model the signal visible at TP7 is a sum between the sampled signal and the
interference. The interference is produced by another oscillator on the workboard (Osc
4) and its amplitude can be adjusted from the knob ‘Output Osc4/Itf’ between 1 and 2 V.
Spectrum analysis model
The spectrum analysis model, presented in Fig. 3.11, adds some improvements
to the model used for the study of aliasing errors. The main improvement is the
availability of a spectrum analyzer as a virtual instrument (the virtual instrument can be
switched between oscilloscope and spectrum analyzer). An additional frequency meter
is connected to the output of the system. The available test points and adjustable settings
are the same as in the case of the aliasing errors model.

3.3 Practice and exercises

Run the basic sampling model.
1. Set the frequency of the input signal (Osc 1) to approximately 800 Hz with an
amplitude of 2 V. Observe the signals in TP1, TP5, TP7 and TP8.
2. Use the oscilloscope (virtual or physical) to measure the period of the sampling
pulses. Calculate the frequency of the pulses (sampling frequency).

20u x 6div= 120u

fs=1/Ts => 1/120u=10^6/120=10^5/12=8333Hz=8,3kHz

3. For each of the following frequencies of the input signal, measure and write down
the amplitude of the reconstructed signal:
fIN [kHz] 0.3 0.8 1 2 2.4 2.5 2.7 3 3.2 3.4 3.8

AOUT [V] 2 2 2 1,6 1.6 1,6 1,2 0,8 0,4 0,2 0,1
Draw the characteristic of the amplitude of the reconstructed signal with respect to
the frequency of the input signal. Explain the variation of the amplitude values.
4. Set the frequency of the input signal to 1 kHz. Measure the amplitude of the
reconstructed signal for each of the available values of the duty cycle for the
sampling pulses. Compare the measured values with the amplitude of the input
signal. Explain the differences.
-duty cycle affects the amplitudes…………………………………………………

5. Measure the gap between consecutive samples (TP7) for each of the available values
of the duty cycle. What usage can be made of these gaps?
-daca ele sunt destul de mari cee ace putem face e sa adaugam un semnal si putem
combina un semnal cu celalalt
36 Sampling of the voice signal for telephony – 3
Run the aliasing errors model.
6. Set the amplitude of the input signal to 2 V. For each of the following frequencies
of the input signal, measure and write down the values of the amplitude and
frequency of the output signal:
fIN 1 2 2.5 3 3.5 4 4.5 5 5.5 6 7
[kHz]
fOUT 1 0,2 2,5 3.03 3,05 3,05 3,05 5 2,5
[kHz] 0,2 3,03
AOUT 0,75 0,70 0,65 0,25 0,2 0,1 0,1 0,5
[V] 0,7 0,65 0,25

Based on the measured values draw the fOUT fIN characteristic. Explain the
A
variations. What about the amplitude ( OUT ) variations?
Run the Interference model
7. Set the amplitude of the interference (Osc 4) to minimum. Observe the reconstructed
signal at the output, while slowly increasing the amplitude of interference (up to the
maximum value). Is the reconstructed signal a good approximation of the input or is
it influenced by the interference? Why?
Run the spectrum analysis model. Set the virtual instrument to spectrum
analyzer.
8. For each of the following values of the frequency of the input signal use the spectrum
analyzer to view the signals at TP1, TP7 and TP8: 2 kHz, 3 kHz, 4kHz, 5 kHz, 6
kHz. Write down the significant frequency components in each case and specify for
the components with frequencies lower than 4 kHz if they represent the useful signal
or aliasing components.

Questions:
1) Which is the standard sampling rate for the telephony voice? For what
reasons was this value chosen?

The standard sampling rate for telephony voice is 8kHz or 8000 samples per
second. This value was chosen for several reasons: bandwidth limitation,
voice frequency range, compatibility, and cost-effectiveness. Overall, the
8kHz sampling rate for telephony voice is a well-established standard that
meets the requirements of voice transmission while also being practical and
cost-effective.

2) How does the duty cycle of the sampling pulses influence the outcome of
the sampling and reconstruction process? Which of the available duty cycles
should be preferred in a telephony application? Why?
3) Above which value of the input frequency the output frequency does no
longer match that at the input?
 37
4) Where is likely for the telephone signal to encounter cross-talk?

Cross-talk can occur at any point in the telephone network where multiple
signals are present in close proximity to each other. To minimize cross-talk,
telephone companies use various technique such as proper cable shielding,
twisted pair wiring, and signal filtering.

5) Give some examples of possible sources of interference for a telephony

transmission.

Some examples: Electromagnetic interference, radio frequency interference,

cross-talk, impulse noise, environmental noise, etc.

6) Calculate the expected frequency of the reconstructed signal if the frequency

of the input signal is 1.2 kHz.
7) Calculate the expected frequency of the reconstructed signal if the frequency
of the input signal is 6.8 kHz. Compare this value with that calculated at 6).

References:
[rs1] Oppenheim, A; Willsky, A; Nawab, H, Signals and Systems 2nd edition,
1996, Prentice-Hall.
[rs2] Feedback Instruments, Teknikit Telephony Training System – Student’s
Workbook 58-001-USB(WB), 2003, FI Ltd. Crowborough UK